Accurate performance of pitch determination algorithms (PDAs) is essential to obtain good quality speech coding with linear prediction at low bit-rates. In this study, five pitch determination algorithms, representative of the range of short-term analysis methods, are applied to the residual from a linear prediction inverse filter. Four of these algorithms are well known (autocorrelation, amdf, cepstrum and maximum likelihood) while the fifth is a novel harmonic analysis technique applying the frequency domain autocorrelation to the power spectrum of the residual. Other comparative studies have generally used added gaussian noise to test robustness. In this study, both additive and multiplicative noise is combined with the speech at various levels and used to test the algorithms. Results indicate that multiplicative noise can have severe consequences on the accuracy of the pitch determination algorithms and that the novel harmonic analysis method performs well under adverse conditions.